inspect-ai 0.3.84__py3-none-any.whl → 0.3.86__py3-none-any.whl

inspect_ai/_eval/eval.py

@@ -43,7 +43,7 @@ from inspect_ai.model import (
     GenerateConfigArgs,
     Model,
 )
-from inspect_ai.model._model import init_active_model, resolve_models
+from inspect_ai.model._model import get_model, init_active_model, resolve_models
 from inspect_ai.scorer._reducer import reducer_log_names
 from inspect_ai.solver._chain import chain
 from inspect_ai.solver._solver import Solver, SolverSpec
@@ -751,10 +751,15 @@ async def eval_retry_async(
             else None
         )
 
+        # resolve the model
+        model = get_model(
+            model=eval_log.eval.model,
+            config=eval_log.eval.model_generate_config,
+            base_url=eval_log.eval.model_base_url,
+            **eval_log.eval.model_args,
+        )
+
         # collect the rest of the params we need for the eval
-        model = eval_log.eval.model
-        model_base_url = eval_log.eval.model_base_url
-        model_args = eval_log.eval.model_args
         task_args = eval_log.eval.task_args
         tags = eval_log.eval.tags
         limit = eval_log.eval.config.limit
@@ -813,8 +818,6 @@ async def eval_retry_async(
                     id=task_id, task=task, task_args=task_args, model=None, log=eval_log
                 ),
                 model=model,
-                model_base_url=model_base_url,
-                model_args=model_args,
                 task_args=task_args,
                 sandbox=eval_log.eval.sandbox,
                 sandbox_cleanup=sandbox_cleanup,

inspect_ai/_eval/task/log.py

@@ -139,6 +139,7 @@ class TaskLogger:
             tags=tags,
             solver_args=solver.args if solver else None,
             model=str(ModelName(model)),
+            model_generate_config=model.config,
             model_base_url=model.api.base_url,
             dataset=EvalDataset(
                 name=dataset.name,

inspect_ai/agent/_agent.py

@@ -225,16 +225,12 @@ def agent_with(
         name = name or info.name
         description = description or info.metadata.get(AGENT_DESCRIPTION, None)
 
-    # if the name is null then raise
-    if name is None:
-        raise ValueError("You must provide a name to agent_with")
-
     # now set registry info
     set_registry_info(
         agent,
         RegistryInfo(
             type="agent",
-            name=name,
+            name=name or "agent",
             metadata={AGENT_DESCRIPTION: description}
             if description is not None
             else {},

inspect_ai/log/_log.py

@@ -599,6 +599,9 @@ class EvalSpec(BaseModel):
     model: str
     """Model used for eval."""
 
+    model_generate_config: GenerateConfig = Field(default_factory=GenerateConfig)
+    """Generate config specified for model instance."""
+
     model_base_url: str | None = Field(default=None)
     """Optional override of model base url"""
 

inspect_ai/log/_recorders/buffer/database.py

@@ -199,28 +199,36 @@ class SampleBufferDatabase(SampleBuffer):
             )
 
     def remove_samples(self, samples: list[tuple[str | int, int]]) -> None:
+        # short circuit no samples
+        if len(samples) == 0:
+            return
+
         with self._get_connection(write=True) as conn:
             cursor = conn.cursor()
             try:
-                # Convert list of tuples into a string for SQL IN clause
-                # Format: (('id1', 1), ('id2', 2))
-                sample_conditions = ",".join(
-                    [f"('{sid}', {epoch})" for sid, epoch in samples]
+                # Build a query using individual column comparisons instead of row values
+                placeholders = " OR ".join(
+                    ["(sample_id=? AND sample_epoch=?)" for _ in samples]
                 )
 
-                # Delete associated events first due to foreign key constraint
+                # Flatten parameters for binding
+                parameters = [item for tup in samples for item in tup]
+
+                # Delete associated events first
                 events_query = f"""
                     DELETE FROM events
-                    WHERE (sample_id, sample_epoch) IN ({sample_conditions})
+                    WHERE {placeholders}
                 """
-                cursor.execute(events_query)
+                cursor.execute(events_query, parameters)
+
+                # Then delete the samples using the same approach
+                placeholders = " OR ".join(["(id=? AND epoch=?)" for _ in samples])
 
-                # Then delete the samples
                 samples_query = f"""
                     DELETE FROM samples
-                    WHERE (id, epoch) IN ({sample_conditions})
+                    WHERE {placeholders}
                 """
-                cursor.execute(samples_query)
+                cursor.execute(samples_query, parameters)
             finally:
                 cursor.close()
 
@@ -259,7 +267,7 @@ class SampleBufferDatabase(SampleBuffer):
 
                 # fetch data
                 return Samples(
-                    samples=list(self._get_samples(conn)),
+                    samples=list(self._get_samples(conn, True)),
                     metrics=task_data.metrics,
                     refresh=self.update_interval,
                     etag=str(task_data.version),

inspect_ai/model/_openai.py

@@ -3,7 +3,7 @@ import re
 from copy import copy
 from typing import Literal
 
-from openai import BadRequestError, OpenAIError
+from openai import APIStatusError, OpenAIError
 from openai.types.chat import (
     ChatCompletion,
     ChatCompletionAssistantMessageParam,
@@ -518,7 +518,7 @@ def chat_choices_from_openai(
 
 
 def openai_handle_bad_request(
-    model_name: str, e: BadRequestError
+    model_name: str, e: APIStatusError
 ) -> ModelOutput | Exception:
     # extract message
     if isinstance(e.body, dict) and "message" in e.body.keys():

inspect_ai/model/_providers/openai.py

@@ -13,6 +13,7 @@ from openai import (
     AsyncOpenAI,
     BadRequestError,
     RateLimitError,
+    UnprocessableEntityError,
 )
 from openai._types import NOT_GIVEN
 from openai.types.chat import ChatCompletion
@@ -295,13 +296,13 @@ class OpenAIAPI(ModelAPI):
                     else None
                 ),
             ), model_call()
-        except BadRequestError as e:
+        except (BadRequestError, UnprocessableEntityError) as e:
             return self.handle_bad_request(e), model_call()
 
     def on_response(self, response: dict[str, Any]) -> None:
         pass
 
-    def handle_bad_request(self, ex: BadRequestError) -> ModelOutput | Exception:
+    def handle_bad_request(self, ex: APIStatusError) -> ModelOutput | Exception:
         return openai_handle_bad_request(self.model_name, ex)
 
     def _chat_choices_from_response(

inspect_ai/model/_providers/providers.py

@@ -253,28 +253,6 @@ def none() -> type[ModelAPI]:
     return NoModel
 
 
-@modelapi("goodfire")
-def goodfire() -> type[ModelAPI]:
-    """Get the Goodfire API provider."""
-    FEATURE = "Goodfire API"
-    PACKAGE = "goodfire"
-    MIN_VERSION = "0.3.4"  # Support for newer Llama models and OpenAI compatibility
-
-    # verify we have the package
-    try:
-        import goodfire  # noqa: F401
-    except ImportError:
-        raise pip_dependency_error(FEATURE, [PACKAGE])
-
-    # verify version
-    verify_required_version(FEATURE, PACKAGE, MIN_VERSION)
-
-    # in the clear
-    from .goodfire import GoodfireAPI
-
-    return GoodfireAPI
-
-
 def validate_openai_client(feature: str) -> None:
     FEATURE = feature
     PACKAGE = "openai"

inspect_ai/model/_providers/together.py

@@ -3,7 +3,7 @@ from json import dumps
 from typing import Any
 
 import httpx
-from openai import BadRequestError
+from openai import APIStatusError
 from openai.types.chat import (
     ChatCompletion,
 )
@@ -105,7 +105,7 @@ class TogetherAIAPI(OpenAIAPI):
         return DEFAULT_MAX_TOKENS
 
     @override
-    def handle_bad_request(self, ex: BadRequestError) -> ModelOutput | Exception:
+    def handle_bad_request(self, ex: APIStatusError) -> ModelOutput | Exception:
         response = ex.response.json()
         if "error" in response and "message" in response.get("error"):
             content = response.get("error").get("message")

{inspect_ai-0.3.84.dist-info → inspect_ai-0.3.86.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: inspect_ai
-Version: 0.3.84
+Version: 0.3.86
 Summary: Framework for large language model evaluations
 Author: UK AI Security Institute
 License: MIT License
@@ -56,7 +56,6 @@ Requires-Dist: aioboto3; extra == "dev"
 Requires-Dist: azure-ai-inference; extra == "dev"
 Requires-Dist: google-cloud-aiplatform; extra == "dev"
 Requires-Dist: google-genai; extra == "dev"
-Requires-Dist: goodfire; extra == "dev"
 Requires-Dist: griffe; extra == "dev"
 Requires-Dist: groq; extra == "dev"
 Requires-Dist: ipython; extra == "dev"

{inspect_ai-0.3.84.dist-info → inspect_ai-0.3.86.dist-info}/RECORD

@@ -46,7 +46,7 @@ inspect_ai/_display/textual/widgets/transcript.py,sha256=zaxlDixT6Fie0acAWBM9Hlt
 inspect_ai/_display/textual/widgets/vscode.py,sha256=YTXdIZ0fcf9XE2v3rWIfUTgnXFww8uKCo7skugQLIbs,1247
 inspect_ai/_eval/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 inspect_ai/_eval/context.py,sha256=gWTjEEMVTJMJpCCKLRs4joZDkG00rzE7-HXZFyzSC_I,1283
-inspect_ai/_eval/eval.py,sha256=3VHOUYhkTmni-AaT4dYE9kfoXME68eCqOv_T6xFySzo,40266
+inspect_ai/_eval/eval.py,sha256=tfIYOJSNGNfZHl18XYqXZHIMKlCmUZ8gbqe7I0OZJII,40307
 inspect_ai/_eval/evalset.py,sha256=FnZBVi5hOt6f84PNYlFhkjb7N1lNgiQydQlernJZeW4,24005
 inspect_ai/_eval/list.py,sha256=VbZ-2EI6MqrXvCN7VTz21TQSoU5K5_Q0hqhxmj5A_m0,3744
 inspect_ai/_eval/loader.py,sha256=yCDrW5MhP6GT329hZ_gUm_eAMsCA9G7jb8sm45Pj-pw,24970
@@ -59,7 +59,7 @@ inspect_ai/_eval/task/epochs.py,sha256=Ci7T6CQniSOTChv5Im2dCdSDrP-5hq19rV6iJ2uBc
 inspect_ai/_eval/task/error.py,sha256=Vhqinfdf0eIrjn7kUY7-id8Kbdggr-fEFpAJeJrkJ1M,1244
 inspect_ai/_eval/task/generate.py,sha256=C9-S9ak4VFQO7QgtUbGjt8F4sTyXS5nekR3Mg_MPwmM,2511
 inspect_ai/_eval/task/images.py,sha256=nTzHizlyuPYumPH7gAOBSrNkTwTbAmZ7tKdzN7d_R2k,4035
-inspect_ai/_eval/task/log.py,sha256=w1Uu3VplvL_UUqyCVDmUMOG5s8_E3si6OkglE7xPxM0,11769
+inspect_ai/_eval/task/log.py,sha256=PD2ZrqtHY0zRyx7pB8L5v-txyaBRePs76cFu5Fb-vjE,11817
 inspect_ai/_eval/task/resolved.py,sha256=OCQc_0HmW_Vw8o1KisX0DCn-eOPkTbR1v_y_jEaAlhU,966
 inspect_ai/_eval/task/results.py,sha256=x4weYRK2XGowfBG3f2msOeZQ_pxh230HTlw6kps33jw,17925
 inspect_ai/_eval/task/run.py,sha256=RS2Qv3AythSkQL4fsgBFaXfyx2WDIZuFj9v6ifoRiYs,38714
@@ -450,7 +450,7 @@ inspect_ai/_view/www/src/workspace/tabs/SamplesTab.tsx,sha256=s6jt1-5_Hrgz3_ysT1
 inspect_ai/_view/www/src/workspace/tabs/grouping.ts,sha256=6lvFzReQKQ_43S20xN4kfBJN2F7Tfs2VWeSMIuHxUAI,6187
 inspect_ai/_view/www/src/workspace/tabs/types.ts,sha256=Wa1Y4tZwYO_QJr0Tg9-5xJFztmcMYCODSm6JvdzMpDw,471
 inspect_ai/agent/__init__.py,sha256=nzL9TPAARSJVZRPogWHxZ-qJriXBGmFUM9DV4NRi21o,749
-inspect_ai/agent/_agent.py,sha256=g0hw6sDTXg_4NRjs5Ohze404HzyyIyFNFDvlgGqL2Vw,7736
+inspect_ai/agent/_agent.py,sha256=5MXMrY5bsBQ4AI5y5rVaIYywp7JxQzUnhp-KqBIMF7I,7622
 inspect_ai/agent/_as_solver.py,sha256=_6H0L9JidC6JjpMaBRBAjIrgzE8GKEoygJjOC_JRoLQ,2340
 inspect_ai/agent/_as_tool.py,sha256=vT5hrcKfkyP90i4Ieuy_dx4cYsFKOMdPs-6x12cuqMk,4449
 inspect_ai/agent/_filter.py,sha256=qnT0HbT4edpDi0MwXY3Q3It2pzNRkTRXZDOqfCwMY6M,1234
@@ -510,7 +510,7 @@ inspect_ai/log/_bundle.py,sha256=5Uy-s64_SFokZ7WRzti9mD7yoKrd2sOzdvqKyahoiC4,804
 inspect_ai/log/_condense.py,sha256=OedMphK5Q2YPuY1cnoAM7tGsyVIU6Kwrv3oIeb3dFmY,10881
 inspect_ai/log/_convert.py,sha256=qn6q10Um2XV7dnK4nQargANa0bz6RFJPmaEMINv38cs,3467
 inspect_ai/log/_file.py,sha256=QjeVUegoCWVUv6CMsj0das_UpZZZMfnbvCQAKlFYGXE,17105
-inspect_ai/log/_log.py,sha256=f4ChtLdNc_z0qVXsJCmZyW6BdbFKGTfHWY5gaymsUkc,24970
+inspect_ai/log/_log.py,sha256=KsssY2kGfuDHGIXOGJHN4bO1LXVs0f3XtqIUfA2R68A,25109
 inspect_ai/log/_message.py,sha256=QofM_JZF_x3k_5ta1uQzoN_VnMoUhXFnqWurIn9FXOY,1999
 inspect_ai/log/_retry.py,sha256=e7a2hjl3Ncl8b8sU7CsDpvK8DV0b1uSRLeokRX1mt34,2109
 inspect_ai/log/_samples.py,sha256=wPQlV1VR9djWaj37lLrjBprCabdAm4S2vFOsQTcd12U,4910
@@ -524,7 +524,7 @@ inspect_ai/log/_recorders/recorder.py,sha256=zDDpl2tktPjb6xk5kd4TyEMxkXZiLgXXpPi
 inspect_ai/log/_recorders/types.py,sha256=Aeo-U7FhmWQSvE_uz3fwUI7cqaSR-ZE_uRVu-1fBCgc,865
 inspect_ai/log/_recorders/buffer/__init__.py,sha256=6DsRdnNl-ic-xJmnBE5i45ZP3eB4yAta9wxi5WFcbqc,367
 inspect_ai/log/_recorders/buffer/buffer.py,sha256=rtLvaX7nSqNrWb-3CeSaOHwJgF1CzRgXFT_I1dDkM1k,945
-inspect_ai/log/_recorders/buffer/database.py,sha256=aqBJdM6meQTWsLs9uF1gFGg1dsE1MvVQdiXR1DHoRqw,22171
+inspect_ai/log/_recorders/buffer/database.py,sha256=3yV8OlDsQ4zFQHNqe7aBAHwkUISW3zmaLBlD1OFj36w,22396
 inspect_ai/log/_recorders/buffer/filestore.py,sha256=S6RP-5zkOPSmy1hV2LCCbfwdX-YFZGuIEjfJuOWMjDQ,8274
 inspect_ai/log/_recorders/buffer/types.py,sha256=pTnPCZHbk9qF6yF-eNXHTa23cLH_FvP8dmfPJCFO15Q,2046
 inspect_ai/model/__init__.py,sha256=6Aa_HEU-rgxWPDaIRlE6KBdXY406x2LtcLeVtAxk-AI,2453
@@ -537,7 +537,7 @@ inspect_ai/model/_generate_config.py,sha256=_-kzw7LOl45baVkTjlfL1K1VLKGgNOOczH2H
 inspect_ai/model/_model.py,sha256=h4ASS2VuTZ_97145rLW202u6e7-mw4ENnnlBl0Vsbio,52127
 inspect_ai/model/_model_call.py,sha256=VJ8wnl9Y81JaiClBYM8eyt1jVb3n-yc6Dd88ofRiJDc,2234
 inspect_ai/model/_model_output.py,sha256=R5EAUPLc5RWymVb3le4cbqbNCZ9voTzg0U1j_e4I-yM,7768
-inspect_ai/model/_openai.py,sha256=0OAmxQbIU6V7WJr9Q8J6oGwQuY9aZLPpHQ9r28GCmbg,19382
+inspect_ai/model/_openai.py,sha256=-N_LhZR8-nrnCL8h9lklo_RrGNDR1SzMJ0tPafVuPXo,19380
 inspect_ai/model/_openai_computer_use.py,sha256=vbKkYLhqNuX16zuWfg5MaGp9H8URrPcLhKQ1pDsZtPo,5943
 inspect_ai/model/_openai_responses.py,sha256=bQWuVvJIkS8CqtoX9z1aRb1aky4TNbMngG2paB3wsrA,20179
 inspect_ai/model/_reasoning.py,sha256=qmR8WT6t_cb7NIsJOQHPyFZh2eLV0HmYxKo2vtvteQ4,929
@@ -547,7 +547,6 @@ inspect_ai/model/_providers/anthropic.py,sha256=PYxV0D_bt0Icp2wEWb6GMCpDb-uBFKYy
 inspect_ai/model/_providers/azureai.py,sha256=uXED_qmeyW1XAGBosbG7PJNk833RIeokKX3l_8O9gYA,14341
 inspect_ai/model/_providers/bedrock.py,sha256=rh8BvSUPWiFMh0TQwMYTlucfFrDKswtLhzozulrz7wE,24004
 inspect_ai/model/_providers/cloudflare.py,sha256=mWqBqc0zzf29UWz34biq8CxSu99a95YjpH_6A4na52g,4617
-inspect_ai/model/_providers/goodfire.py,sha256=J0nxGbF8lXBmc5YHBJCsZdF03mWT5SuWMb21d9ho3FM,8799
 inspect_ai/model/_providers/google.py,sha256=gcg8pvYAV5gYc4NXC5mLqFyuU7KuhyNrzdXIY57sYl8,28207
 inspect_ai/model/_providers/grok.py,sha256=dS88ueXiD-kHAFr0jCoTpTGLGa2VsUlB_TFP8L_2lBM,995
 inspect_ai/model/_providers/groq.py,sha256=mcRKu33e-mO5l06PGV6SjsildQd0XCti6QNXwwFWL7I,11246
@@ -557,12 +556,12 @@ inspect_ai/model/_providers/mistral.py,sha256=FbMPN_pw8LZal2iFGf5FX70ypuH3k44FUn
 inspect_ai/model/_providers/mockllm.py,sha256=gL9f-f5TOdE4a0GVENr3cOIIp2kv8zVXWPZ608rouGk,2440
 inspect_ai/model/_providers/none.py,sha256=6qLbZpHSoEZaaxFO7luieFjqig2Ju8Fu00DlRngAry8,935
 inspect_ai/model/_providers/ollama.py,sha256=mBPSxaEkiH_RnlHKqOyFBlXObQhc2dfjL-rCKrea5u8,675
-inspect_ai/model/_providers/openai.py,sha256=NFdMpnI2vlmpI8h_vWnt8y4X_XaydaL9gH5Dmy6k5Tw,16478
+inspect_ai/model/_providers/openai.py,sha256=zJkhtiEQrmsuhfL7mpBPpOlYJ_WNraeyTkjYTelF0no,16535
 inspect_ai/model/_providers/openai_o1.py,sha256=k-Xm_Wzn1KHKL6Z1KTHg4CTTr8ybgiHvXkLiLdjP7Os,12926
 inspect_ai/model/_providers/openai_responses.py,sha256=YPXt8KQfIEiiTpvtoQECBoNQLDLbwBW_KhBfM8vEhJk,6324
 inspect_ai/model/_providers/openrouter.py,sha256=pDimDmm_4FzS4GZx0n9z8z717mQf3IQlgEy30huzpc4,4730
-inspect_ai/model/_providers/providers.py,sha256=0WSi_FOWxW71sZ4GJ-OgJqbPS4tMIaPQqEG2hnxqfqc,6378
-inspect_ai/model/_providers/together.py,sha256=MoA3tyMKUnE0EekTqEIBBwvsaOp5c697kydLi1ZMYzE,9745
+inspect_ai/model/_providers/providers.py,sha256=Sd2D9OcWkukuBcl_-KDfdpxMaAShv1JZhL5KfAM87CE,5817
+inspect_ai/model/_providers/together.py,sha256=Wh3G0vhKHq5ofx1otwXjJFhM98Ll70IbqBhUNNV2-rk,9743
 inspect_ai/model/_providers/vertex.py,sha256=60W7kgoA83GtKdMeJgNU2IAw0N0wTscg4YCcMPu2bwo,17185
 inspect_ai/model/_providers/vllm.py,sha256=UYjCCXzw2hGJHVC3oPl-u2EI4iAm8ZncoIfYp1QJkbQ,14238
 inspect_ai/model/_providers/util/__init__.py,sha256=d4T_qvXihTRd1zmQkNE3xUBlHCX8tOIbRK19EwU0fTs,717
@@ -693,9 +692,9 @@ inspect_ai/util/_sandbox/docker/internal.py,sha256=c8X8TLrBPOvsfnq5TkMlb_bzTALyc
 inspect_ai/util/_sandbox/docker/prereqs.py,sha256=0j6_OauBBnVlpBleADcZavIAAQZy4WewVjbRn9c0stg,3355
 inspect_ai/util/_sandbox/docker/service.py,sha256=hhHIWH1VDFLwehdGd19aUBD_VKfDO3GCPxpw1HSwVQk,2437
 inspect_ai/util/_sandbox/docker/util.py,sha256=EeInihCNXgUWxaqZ4dNOJd719kXL2_jr63QCoXn68vA,3154
-inspect_ai-0.3.84.dist-info/licenses/LICENSE,sha256=xZPCr8gTiFIerrA_DRpLAbw-UUftnLFsHxKeW-NTtq8,1081
-inspect_ai-0.3.84.dist-info/METADATA,sha256=g-2UAMeNEN0cyQB6JUowoPVFebPlFpNsZuFiQwsxpVE,5005
-inspect_ai-0.3.84.dist-info/WHEEL,sha256=CmyFI0kx5cdEMTLiONQRbGQwjIoR1aIYB7eCAQ4KPJ0,91
-inspect_ai-0.3.84.dist-info/entry_points.txt,sha256=WGGLmzTzDWLzYfiyovSY6oEKuf-gqzSDNOb5V-hk3fM,54
-inspect_ai-0.3.84.dist-info/top_level.txt,sha256=Tp3za30CHXJEKLk8xLe9qGsW4pBzJpEIOMHOHNCXiVo,11
-inspect_ai-0.3.84.dist-info/RECORD,,
+inspect_ai-0.3.86.dist-info/licenses/LICENSE,sha256=xZPCr8gTiFIerrA_DRpLAbw-UUftnLFsHxKeW-NTtq8,1081
+inspect_ai-0.3.86.dist-info/METADATA,sha256=tId3lj5ywe2A79iWACtXXP_aBeWfnbQQfAujxJGxaoc,4965
+inspect_ai-0.3.86.dist-info/WHEEL,sha256=CmyFI0kx5cdEMTLiONQRbGQwjIoR1aIYB7eCAQ4KPJ0,91
+inspect_ai-0.3.86.dist-info/entry_points.txt,sha256=WGGLmzTzDWLzYfiyovSY6oEKuf-gqzSDNOb5V-hk3fM,54
+inspect_ai-0.3.86.dist-info/top_level.txt,sha256=Tp3za30CHXJEKLk8xLe9qGsW4pBzJpEIOMHOHNCXiVo,11
+inspect_ai-0.3.86.dist-info/RECORD,,

inspect_ai/model/_providers/goodfire.py

@@ -1,253 +0,0 @@
-import os
-from typing import Any, List, Literal, get_args
-
-from goodfire import AsyncClient
-from goodfire.api.chat.interfaces import ChatMessage as GoodfireChatMessage
-from goodfire.api.exceptions import (
-    InvalidRequestException,
-    RateLimitException,
-    ServerErrorException,
-)
-from goodfire.variants.variants import SUPPORTED_MODELS, Variant
-from typing_extensions import override
-
-from inspect_ai.tool._tool_choice import ToolChoice
-from inspect_ai.tool._tool_info import ToolInfo
-
-from .._chat_message import (
-    ChatMessage,
-    ChatMessageAssistant,
-    ChatMessageSystem,
-    ChatMessageTool,
-    ChatMessageUser,
-)
-from .._generate_config import GenerateConfig
-from .._model import ModelAPI
-from .._model_call import ModelCall
-from .._model_output import (
-    ChatCompletionChoice,
-    ModelOutput,
-    ModelUsage,
-)
-from .util import environment_prerequisite_error, model_base_url
-
-# Constants
-GOODFIRE_API_KEY = "GOODFIRE_API_KEY"
-DEFAULT_BASE_URL = "https://api.goodfire.ai"
-DEFAULT_MAX_TOKENS = 4096
-DEFAULT_TEMPERATURE = 1.0  # Standard sampling temperature (baseline)
-DEFAULT_TOP_P = 1.0  # No nucleus sampling truncation (baseline)
-
-
-class GoodfireAPI(ModelAPI):
-    """Goodfire API provider.
-
-    This provider implements the Goodfire API for LLM inference. It supports:
-    - Chat completions with standard message formats
-    - Basic parameter controls (temperature, top_p, etc.)
-    - Usage statistics tracking
-    - Stop reason handling
-
-    Does not currently support:
-    - Tool calls
-    - Feature analysis
-    - Streaming responses
-
-    Known limitations:
-    - Limited role support (system/user/assistant only)
-    - Tool messages converted to user messages
-    """
-
-    client: AsyncClient
-    variant: Variant
-    model_args: dict[str, Any]
-
-    def __init__(
-        self,
-        model_name: str,
-        base_url: str | None = None,
-        api_key: str | None = None,
-        config: GenerateConfig = GenerateConfig(),
-        **model_args: Any,
-    ) -> None:
-        """Initialize the Goodfire API provider.
-
-        Args:
-            model_name: Name of the model to use
-            base_url: Optional custom API base URL
-            api_key: Optional API key (will check env vars if not provided)
-            config: Generation config options
-            **model_args: Additional arguments passed to the API
-        """
-        super().__init__(
-            model_name=model_name,
-            base_url=base_url,
-            api_key=api_key,
-            api_key_vars=[GOODFIRE_API_KEY],
-            config=config,
-        )
-
-        # resolve api_key
-        if not self.api_key:
-            self.api_key = os.environ.get(GOODFIRE_API_KEY)
-            if not self.api_key:
-                raise environment_prerequisite_error("Goodfire", GOODFIRE_API_KEY)
-
-        # Validate model name against supported models
-        supported_models = list(get_args(SUPPORTED_MODELS))
-        if self.model_name not in supported_models:
-            raise ValueError(
-                f"Model {self.model_name} not supported. Supported models: {supported_models}"
-            )
-
-        # Initialize client with minimal configuration
-        base_url_val = model_base_url(base_url, "GOODFIRE_BASE_URL")
-        assert isinstance(base_url_val, str) or base_url_val is None
-
-        # Store model args for use in generate
-        self.model_args = model_args
-
-        self.client = AsyncClient(
-            api_key=self.api_key,
-            base_url=base_url_val or DEFAULT_BASE_URL,
-        )
-
-        # Initialize variant directly with model name
-        self.variant = Variant(self.model_name)  # type: ignore
-
-    def _to_goodfire_message(self, message: ChatMessage) -> GoodfireChatMessage:
-        """Convert an Inspect message to a Goodfire message format.
-
-        Args:
-            message: The message to convert
-
-        Returns:
-            The converted message in Goodfire format
-
-        Raises:
-            ValueError: If the message type is unknown
-        """
-        role: Literal["system", "user", "assistant"] = "user"
-        if isinstance(message, ChatMessageSystem):
-            role = "system"
-        elif isinstance(message, ChatMessageUser):
-            role = "user"
-        elif isinstance(message, ChatMessageAssistant):
-            role = "assistant"
-        elif isinstance(message, ChatMessageTool):
-            role = "user"  # Convert tool messages to user messages
-        else:
-            raise ValueError(f"Unknown message type: {type(message)}")
-
-        content = str(message.content)
-        if isinstance(message, ChatMessageTool):
-            content = f"Tool {message.function}: {content}"
-
-        return GoodfireChatMessage(role=role, content=content)
-
-    def handle_error(self, ex: Exception) -> ModelOutput | Exception:
-        """Handle only errors that need special treatment for retry logic or model limits."""
-        # Handle token/context length errors
-        if isinstance(ex, InvalidRequestException):
-            error_msg = str(ex).lower()
-            if "context length" in error_msg or "max tokens" in error_msg:
-                return ModelOutput.from_content(
-                    model=self.model_name,
-                    content=str(ex),
-                    stop_reason="model_length",
-                    error=error_msg,
-                )
-
-        # Let all other errors propagate
-        return ex
-
-    @override
-    def should_retry(self, ex: Exception) -> bool:
-        """Check if exception is due to rate limiting."""
-        return isinstance(ex, RateLimitException | ServerErrorException)
-
-    @override
-    def connection_key(self) -> str:
-        """Return key for connection pooling."""
-        return f"goodfire:{self.api_key}"
-
-    @override
-    def max_tokens(self) -> int | None:
-        """Return maximum tokens supported by model."""
-        return DEFAULT_MAX_TOKENS  # Let Goodfire's Variant handle model-specific limits
-
-    async def generate(
-        self,
-        input: List[ChatMessage],
-        tools: List[ToolInfo],
-        tool_choice: ToolChoice,
-        config: GenerateConfig,
-        *,
-        cache: bool = True,
-    ) -> tuple[ModelOutput | Exception, ModelCall]:
-        """Generate output from the model."""
-        # Convert messages and prepare request params
-        messages = [self._to_goodfire_message(msg) for msg in input]
-        # Build request parameters with type hints
-        params: dict[str, Any] = {
-            "model": self.variant.base_model,  # Use base_model instead of stringifying the Variant
-            "messages": messages,
-            "max_completion_tokens": int(config.max_tokens)
-            if config.max_tokens
-            else DEFAULT_MAX_TOKENS,
-            "stream": False,
-        }
-
-        # Add generation parameters from config if not in model_args
-        if "temperature" not in self.model_args and config.temperature is not None:
-            params["temperature"] = float(config.temperature)
-        elif "temperature" not in self.model_args:
-            params["temperature"] = DEFAULT_TEMPERATURE
-
-        if "top_p" not in self.model_args and config.top_p is not None:
-            params["top_p"] = float(config.top_p)
-        elif "top_p" not in self.model_args:
-            params["top_p"] = DEFAULT_TOP_P
-
-        # Add any additional model args (highest priority)
-        api_params = {
-            k: v
-            for k, v in self.model_args.items()
-            if k not in ["api_key", "base_url", "model_args"]
-        }
-        params.update(api_params)
-
-        try:
-            # Use native async client
-            response = await self.client.chat.completions.create(**params)
-            response_dict = response.model_dump()
-
-            output = ModelOutput(
-                model=self.model_name,
-                choices=[
-                    ChatCompletionChoice(
-                        message=ChatMessageAssistant(
-                            content=response_dict["choices"][0]["message"]["content"],
-                            model=self.model_name,
-                        ),
-                        stop_reason="stop",
-                    )
-                ],
-                usage=ModelUsage(**response_dict["usage"])
-                if "usage" in response_dict
-                else None,
-            )
-            model_call = ModelCall.create(request=params, response=response_dict)
-            return (output, model_call)
-        except Exception as ex:
-            result = self.handle_error(ex)
-            model_call = ModelCall.create(
-                request=params,
-                response={},  # Empty response for error case
-            )
-            return (result, model_call)
-
-    @property
-    def name(self) -> str:
-        """Get provider name."""
-        return "goodfire"